By Peyman Sabri · Last updated 2026-10-31
Yes, if it could pass as real. From 2 August 2026, Article 50 of the EU AI Act requires a business that publishes AI-generated or manipulated video resembling real people, places or events to disclose it clearly, at the latest when someone first sees it. AI used only as an editing aid, and work that is obviously fictional, is treated far more lightly.
I’m Peyman Sabri, a Stockholm-based content producer and motion designer at Peyman Sabri Media. AI video is part of my everyday work, next to event films, motion design and branding. This is how I read the rules, and how I apply them before anything is published.
This post is not legal advice. For your specific case, read the sources linked below or ask a lawyer.
What does the EU AI Act actually say about AI video?
Article 50 is the transparency part of the AI Act. It has four obligations, and two of them matter for video.
- Article 50(2): the companies that build generative AI tools must mark the output in a machine-readable way, so it can be detected as AI-generated or manipulated.
- Article 50(4): the business that uses the tool and publishes the result must disclose deepfakes to the people who see them.
The other two cover chatbots that talk directly to people, and emotion recognition. Article 50(4) also covers AI-written text on matters of public interest.
According to the European Commission’s FAQ on Article 50, the rules apply from 2 August 2026. There is one limited grace period. AI systems already on the market before that date have until 2 December 2026 to add machine-readable marking. Content published before 2 August 2026 does not have to be labelled after the fact, although the Commission encourages it.
Timeline
- Before 2 Aug 2026Published content does not need a retroactive label.
- 2 Aug 2026Article 50 applies. Deepfakes must be disclosed at first exposure.
- 2 Dec 2026End of the grace period for machine-readable marking in AI tools already on the market.
Am I a provider or a deployer?
Most businesses that make or buy video are deployers. The AI Act defines a deployer as someone using an AI system under their own authority, except for purely personal use.
The provider is the company that builds the tool and puts it on the market. Its job is the invisible part: watermarks and metadata that machines can read.
Your job as a deployer is the visible part. The Commission is clear that a deployer cannot rely on the provider’s machine-readable marking to meet the disclosure duty. People have to be able to see or hear the label without special tools.
Two details are easy to miss:
- A company stays the deployer even when a contractor or freelancer runs the AI tool on its behalf and under its control.
- A freelancer who uses AI professionally is a deployer too.
In practice, agree before production who decides on labelling.
When does an AI video count as a deepfake?
The AI Act’s definition in Article 3(60) is narrower than the everyday word. The Commission breaks it into three conditions that must all be met:
- Resemblance: the content looks or sounds very much like its subject.
- Existing: the subject is a person, object, place, entity or event that exists, or plausibly could.
- False appearance: it would falsely appear to a person to be authentic or truthful.
Does this video need a label?
- Did AI generate or substantially change the image, sound or video? If no, the deepfake rule does not apply.
- Does it resemble a real person, place, object or event? If no, it is not a deepfake.
- Could your audience take it as real? If yes to all three, disclose it clearly, at the latest at first exposure.
The third condition depends on context. The Commission’s guidance lets you weigh the level of resemblance, the message, where the content will be shown and what the audience expects.
A generated voice that sounds like your CEO, in a video on your website, is a clear case. A stylised animated character that resembles no one is not.
What doesn’t need a label?
Three kinds of use are treated lightly or fall outside the rule.
Standard editing. The marking duty does not apply when AI performs an assistive function for standard editing or does not substantially alter the input or its meaning. Think noise reduction or colour matching.
Production work behind the scenes. The Commission’s FAQ gives the example of AI-made background scenes, special effects and technical pre- and post-processing in standard film production. Those are not likely to make content look falsely authentic.
Obviously creative or fictional work. For deepfakes in evidently artistic, creative, satirical or fictional work, the duty is lighter. You still disclose, but in a way that does not hamper the display or enjoyment of the work. How and where is a judgement call, for example a line in the credits or the description.
The grey zone is everything in between. That is where judgement matters more than the tool.
What does a good AI label look like?
Article 50(5) says the information must be clear and distinguishable, given at the latest at the first interaction or exposure, and meet accessibility requirements. PTS summarises it in Swedish as clear, easy to notice and given no later than first contact.
For video, that usually means one or more of these:
- An on-screen label in the first seconds, readable on a phone and long enough to read.
- A spoken line when the AI element is a voice.
- A line in the post text or video description, placed above the fold, not after “see more”.
- Captions and alt text that say the same, so the label works without sound and with a screen reader.
There is also a Code of Practice on transparency of AI-generated content, which the Commission and the AI Board have assessed as adequate. It is voluntary, but businesses that sign it can rely on it to show they comply.
A label done well does not weaken a film. It tells the audience you respect them, and that is part of the message too.
How do I handle AI in my own productions?
Judgement first, tools second. Before I generate a single frame, I ask the same three questions as in the box above, together with whoever owns the film.
Then I write it down: what was generated, what it resembles, and where the label goes. The note travels with the film, so whoever reuses a clip knows what is in it.
I keep real people real. If a person appears to say something, they said it, unless the film is obviously fiction and labelled.
And I treat the rules as a floor. An audience that finds out afterwards that something was generated tends to remember that, not the message.
Frequently asked questions
Does AI-generated video have to be labelled in Sweden?
Yes, when it is a deepfake. The EU AI Act applies in Sweden, and from 2 August 2026 businesses must disclose AI video that could falsely appear real. PTS has published guidance in Swedish.
When do the AI Act’s transparency rules start to apply?
From 2 August 2026. AI tools already on the market get until 2 December 2026 for machine-readable marking. Older content needs no retroactive label.
What counts as a deepfake under the AI Act?
AI-generated or manipulated image, audio or video that resembles real persons, objects, places, entities or events and would falsely appear authentic. All three conditions must be met.
Does a commercial need a label if AI was only used in the edit?
Usually not. AI that only assists standard editing, without substantially changing the content or its meaning, falls outside the marking duty.
Who is responsible for the label, the tool or the business that publishes?
Both, for different things. The tool provider marks the output for machines. The business publishing the video discloses deepfakes to people, even when a freelancer ran the tool.
If you are planning a film with AI in it and want to talk through where the line goes, write to me at hello@peymansabri.com.
